Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory

Researchers say many AI agent-memory systems act like notebooks: they retrieve facts only when a query looks directly relevant.

That creates a blind spot for safety-critical, context-dependent links, such as allergy warnings for macarons or lilies around cats.

Several state-of-the-art systems can answer direct memory questions but still fail to use the same facts through indirect world-knowledge connections.

To measure this weakness, the authors introduce InMind Memory Bench, a benchmark for implicit-association failures.